東華大學圖書館 |

Mangrove Species Classification and Leaf Area Index Estimation from Multispectral, Hyperspectral and LiDAR Data in Mai Po Nature Reserve, Hong Kong.

紀錄類型:	書目-電子資源 : Monograph/item
正題名/作者:	Mangrove Species Classification and Leaf Area Index Estimation from Multispectral, Hyperspectral and LiDAR Data in Mai Po Nature Reserve, Hong Kong./
作者:	Li, Qiaosi.
面頁冊數:	1 online resource (212 pages)
附註:	Source: Dissertations Abstracts International, Volume: 83-01, Section: B.
Contained By:	Dissertations Abstracts International83-01B.
標題:	Geography. -
電子資源:	http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28663006click for full text (PQDT)
ISBN:	9798505585047

Mangrove Species Classification and Leaf Area Index Estimation from Multispectral, Hyperspectral and LiDAR Data in Mai Po Nature Reserve, Hong Kong.
Li, Qiaosi.

Mangrove Species Classification and Leaf Area Index Estimation from Multispectral, Hyperspectral and LiDAR Data in Mai Po Nature Reserve, Hong Kong. - 1 online resource (212 pages)

Source: Dissertations Abstracts International, Volume: 83-01, Section: B.

Thesis (Ph.D.)--The Chinese University of Hong Kong (Hong Kong), 2020.

Includes bibliographical references

Mangroves have significant social, economic, environmental, and ecological values but they are under threat of deforestation due to human activities and climate change. Hong Kong has 60 mangrove stands covering an area of about 510 ha with eight local mangrove species. Mai Po Inner Deep Bay Ramsar Site has the largest mangrove stand in Hong Kong where is categorized as an extremely important stand that must be conserved immediately for its biodiversity. Therefore, understanding mangrove species distribution and monitoring growth and infestation are required for mangrove conservation. With the rapid development recently, remote sensing being a cost-effective and timely tool is the most appropriate technique to this end. Taking Mai Po as the study area, this study explored the potential of spaceborne multispectral data, and airborne hyperspectral (HSI) as well as LiDAR data for mangrove species classification and leaf area index (LAI) estimation. A comprehensive species classification was accomplished by understory detection and upper canopy species classification. First, LiDAR returns were classified into the upper canopy and understory by analyzing the vertical distribution of LiDAR returns, so the understory distribution could be mapped. After that, important features from multispectral data WorldView-3 (WV-3), HSI, and LiDAR data were selected by feature selection method recursive feature elimination based on random forest (RFE-RF). As a consequence, seven feature sets including four single source features and three combination features were obtained. Seven feature sets were input to Random Forest (RF), Support Vector Machine (SVM) and Convolutional Neural Network (CNN) classifiers to compare the performance of different data and classification algorithms. Vegetation indices (VIs) from multispectral data Sentinel 2 (S2) and HSI, as well as LiDAR metrics were derived to predict the LAI. Effective overstory LAI (Leo), understory LAI (Leu), and total LAI (Let) were estimated through parametric and non-parametric regression methods. The impact of understory and gap fraction on LAI estimation was quantitatively analyzed. Important features that were indicated by the feature selection results included seven WV-3 bands (blue, green, yellow, red, red edge, near infrared (NIR) 1, NIR2) and twenty HSI wavelengths (green region of 522 nm, red edge region between 682 - 706 nm and 720 - 744nm, NIR region 784 nm, 873 - 902 nm, and 929 nm). Meanwhile, important LiDAR metrics mainly described canopy heights, distribution of return height (e.g., standard deviation and skewness of height), canopy shape (e.g., canopy relief ratio), and 5th- 20th percentile of the height distribution. LiDAR features were considered more important than spectral features. The Classification accuracy varied from different data and algorithms. CNN algorithm (accuracy: 0.815 - 0.820) outperformed SVM and RF (accuracy: 0.631 - 0.647) when dealt with spectral features. LiDAR features obtained stable accuracy (0.780 - 0.800) across different algorithms. Combining spectral features and LiDAR features greatly improved the classification with accuracy from 0.836 - 0.893. CNN algorithms coupled with combination data generated the best accuracy (0.875 - 0.893) while the RF algorithm was the most efficient with satisfactory accuracy (0.838 - 0.870). VIs that were sensitive to LAI were mainly constructed by red edge, NIR and shortwave infrared (SWIR) bands. LiDAR metrics describing standard variation and skewness of return height, vertical penetration, strata information were sensitive to LAI. S2 VIs performed better to predict Let while HSI VIs were better at Leo estimation. LiDAR metrics were more capable than spectral VIs to predict Leu. Spectral VIs and LiDAR metrics could be complementary to enhance Leo and Leu. Leo could be the most accurately predicted by HSI VIs + LiDAR metrics with RMSE = 0.151 and R2 = 0.802. Leu could be best estimated by S2 VIs + LiDAR metrics with RMSE = 0.424 and R2 = 0.728. Using S2 VI alone could best estimate Let with RMSE = 0.592 and R2 = 0.675. Moreover, Let model fitting was greatly strengthened by considering the effect of gap fraction, which indicated the understory and the gap fraction could significantly affect the canopy spectral response in an optical data. Non-parametric regression often yielded the best accuracy and coped well with combination data.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2023

Mode of access: World Wide Web

ISBN: 9798505585047Subjects--Topical Terms:

524010
Geography.
Subjects--Index Terms:

MangrovesIndex Terms--Genre/Form:

542853
Electronic books.

Mangrove Species Classification and Leaf Area Index Estimation from Multispectral, Hyperspectral and LiDAR Data in Mai Po Nature Reserve, Hong Kong.
LDR:10244nmm a2200421K 4500 001 2365163
005 20231213130601.5
006 m o d
007 cr mn ---uuuuu
008 241011s2020 xx obm 000 0 eng d
020 $a 9798505585047
035 $a (MiAaPQ)AAI28663006
035 $a AAI28663006
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Li, Qiaosi. $3 3705991
245 1 0 $a Mangrove Species Classification and Leaf Area Index Estimation from Multispectral, Hyperspectral and LiDAR Data in Mai Po Nature Reserve, Hong Kong.
264 0 $c 2020
300 $a 1 online resource (212 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 83-01, Section: B.
500 $a Advisor: Kit, Wong Kwan.
502 $a Thesis (Ph.D.)--The Chinese University of Hong Kong (Hong Kong), 2020.
504 $a Includes bibliographical references
520 $a Mangroves have significant social, economic, environmental, and ecological values but they are under threat of deforestation due to human activities and climate change. Hong Kong has 60 mangrove stands covering an area of about 510 ha with eight local mangrove species. Mai Po Inner Deep Bay Ramsar Site has the largest mangrove stand in Hong Kong where is categorized as an extremely important stand that must be conserved immediately for its biodiversity. Therefore, understanding mangrove species distribution and monitoring growth and infestation are required for mangrove conservation. With the rapid development recently, remote sensing being a cost-effective and timely tool is the most appropriate technique to this end. Taking Mai Po as the study area, this study explored the potential of spaceborne multispectral data, and airborne hyperspectral (HSI) as well as LiDAR data for mangrove species classification and leaf area index (LAI) estimation. A comprehensive species classification was accomplished by understory detection and upper canopy species classification. First, LiDAR returns were classified into the upper canopy and understory by analyzing the vertical distribution of LiDAR returns, so the understory distribution could be mapped. After that, important features from multispectral data WorldView-3 (WV-3), HSI, and LiDAR data were selected by feature selection method recursive feature elimination based on random forest (RFE-RF). As a consequence, seven feature sets including four single source features and three combination features were obtained. Seven feature sets were input to Random Forest (RF), Support Vector Machine (SVM) and Convolutional Neural Network (CNN) classifiers to compare the performance of different data and classification algorithms. Vegetation indices (VIs) from multispectral data Sentinel 2 (S2) and HSI, as well as LiDAR metrics were derived to predict the LAI. Effective overstory LAI (Leo), understory LAI (Leu), and total LAI (Let) were estimated through parametric and non-parametric regression methods. The impact of understory and gap fraction on LAI estimation was quantitatively analyzed. Important features that were indicated by the feature selection results included seven WV-3 bands (blue, green, yellow, red, red edge, near infrared (NIR) 1, NIR2) and twenty HSI wavelengths (green region of 522 nm, red edge region between 682 - 706 nm and 720 - 744nm, NIR region 784 nm, 873 - 902 nm, and 929 nm). Meanwhile, important LiDAR metrics mainly described canopy heights, distribution of return height (e.g., standard deviation and skewness of height), canopy shape (e.g., canopy relief ratio), and 5th- 20th percentile of the height distribution. LiDAR features were considered more important than spectral features. The Classification accuracy varied from different data and algorithms. CNN algorithm (accuracy: 0.815 - 0.820) outperformed SVM and RF (accuracy: 0.631 - 0.647) when dealt with spectral features. LiDAR features obtained stable accuracy (0.780 - 0.800) across different algorithms. Combining spectral features and LiDAR features greatly improved the classification with accuracy from 0.836 - 0.893. CNN algorithms coupled with combination data generated the best accuracy (0.875 - 0.893) while the RF algorithm was the most efficient with satisfactory accuracy (0.838 - 0.870). VIs that were sensitive to LAI were mainly constructed by red edge, NIR and shortwave infrared (SWIR) bands. LiDAR metrics describing standard variation and skewness of return height, vertical penetration, strata information were sensitive to LAI. S2 VIs performed better to predict Let while HSI VIs were better at Leo estimation. LiDAR metrics were more capable than spectral VIs to predict Leu. Spectral VIs and LiDAR metrics could be complementary to enhance Leo and Leu. Leo could be the most accurately predicted by HSI VIs + LiDAR metrics with RMSE = 0.151 and R2 = 0.802. Leu could be best estimated by S2 VIs + LiDAR metrics with RMSE = 0.424 and R2 = 0.728. Using S2 VI alone could best estimate Let with RMSE = 0.592 and R2 = 0.675. Moreover, Let model fitting was greatly strengthened by considering the effect of gap fraction, which indicated the understory and the gap fraction could significantly affect the canopy spectral response in an optical data. Non-parametric regression often yielded the best accuracy and coped well with combination data.
520 $a 紅樹林擁有非常重要的社會，經濟，環境及生態價值，但是人類的活動和氣候變化讓它們受到毀林的威脅。香港擁有60 個紅樹林林分，它們總共覆蓋大約 510 公頃，其中包含八種本地紅樹。米埔內后海灣為拉姆薩爾濕地擁有香港面積最大的紅樹林林分，它被視爲極爲重要的林分，需要即刻保育其珍貴的生物多樣性。因此，瞭解紅樹林的樹種分佈和監測其生長狀況及病蟲侵害是紅樹林保育中的急需工作。隨著近年遙感技術的快速發展，業已成爲成本低且能夠高動態監測的技術，是實現香港紅樹林普查和動態監測最為合適的手段。以米埔作爲研究區，本研究實驗並總結了基於星載多光譜數據，機載高光譜 (HSI) 和激光雷達 (LiDAR) 數據在紅樹林樹種分類和葉面積指數 (LAI) 估計的潛力。此研究設計和實現了一項較以往更全面的基於上下層植被數據的樹種分類方法。首先，通過分析激光雷達回波信號的垂直分佈將上層植被和下層植被的激光雷達點雲加以區分，從而獲得下層植被的分佈情況。隨後，利用基於隨機森林的特徵遞歸消減算法從多光譜 WorldView-3 (WV-3)、高光譜和激光雷達數據中選擇重要的特徵。經過特徵選擇最終得到七組特徵子集，包括四組單數據源的子集和三組多數據源的子集。七組特徵子集輸入到隨機森林 (RF)、支持向量機 (SVM) 和卷積神經網絡 (CNN) 分類器中以比較不同數據和算法在樹種分類上的表現。由多光譜數據哨兵2 號 (S2) 和高光譜數據計算而得的植被指數以及激光雷達度量則用於估計葉面積指數 (LAI) 。利用參數和非參數的回歸方法分別估計上層植被有效葉面積指數 (Leo)、下層植被有效葉面積指數 (Leu) 和總有效葉面積指數 (Let) 。下層植被和間隙率對葉面積指數估計的影響也將被定量分析。特徵選擇的結果揭示了一系列重要特徵，它們包括七個WV-3 波段（藍、綠、黃、紅、紅邊、近紅外1 和近紅外2）和二十個高光譜波長（綠波段522 nm，紅邊波段682 nm - 706 nm 及720 nm - 744 nm，近紅外784 nm， 873 nm - 902 nm 及 929 nm）。同時，所選擇的重要激光雷達度量主要描述了樹冠高度，回波信號的高度分佈（例如：點雲高度的標準差和偏斜度），樹冠形狀（例如：樹冠緩率）和第5 至第20 的點雲高度百分位。在樹種分類中激光雷達特徵被認為比光譜特徵更爲重要。分類精度因運用不同的數據和分類算法而異。當僅使用光譜特徵時， CNN 算法的總體精度（0.815 - 0.820）高於SVM 和RF 算法的精度（0.631 - 0.647），而激光雷達特徵在各個算法中獲取相似的精度（0.780 - 0.800）。結合光譜特徵和激光雷達特徵極大地提高了樹種的分類精度（0.836 - 0.893），其中CNN 算法獲得最高的精度（0.875 - 0.893），而RF 算法效率最高並且可獲得較高的精度（0.838 - 0.870）。對葉面積指數敏感的植被指數多由紅邊、近紅外和短波紅外 (SWIR) 構成。對葉面積指數敏感的激光雷達度量主要與點雲高度的標準差和偏斜度、激光脈衝的垂直穿透、植被分層的信息相關。實驗結果表明，S2 的植被指數在對Let 的預測中表現較好，HSI 的植被指數在對Leo 的預測中表現較好，而激光雷達度量比光譜植被指數更適合于預測Leu。光譜植被指數與激光雷達度量優勢互補從而提高了 Leo 和Leu 的精度。結合HSI 植被指數與激光雷達度量為Leo 的預測取得最高的精度，達到RMSE = 0.151 且R2 = 0.802；結合S2 植被指數與激光雷達度量為Leu 的預測取得最高的精度，達到RMSE = 0.424 and R2 = 0.728. 僅使用S2 植被指數即取得最高的Let 預測精度，達到RMSE = 0.592 and R2 = 0.675。此外，當考慮間隙率對計算縂有效葉面積指數的影響時，模型的擬合顯著增强，這表明下層植被和間隙率對光學數據中冠層的光譜響應有著顯著的影響。對於回歸方法而言，非參數回歸常常獲得最高的精度同時也能更好地處理多源結合的數據。
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2023
538 $a Mode of access: World Wide Web
650 4 $a Geography. $3 524010
650 4 $a Natural resource management. $3 589570
650 4 $a Climate change. $2 bicssc $3 2079509
650 4 $a Accuracy. $3 3559958
650 4 $a Regression analysis. $3 529831
650 4 $a Flowers & plants. $3 3564028
650 4 $a Forestry. $3 895157
650 4 $a Discriminant analysis. $3 560438
650 4 $a Data processing. $3 680224
650 4 $a Data analysis. $2 bisacsh $3 3515250
650 4 $a Biomass. $3 1013462
650 4 $a Habitats. $3 3564192
650 4 $a Radiation. $3 673904
650 4 $a Coasts. $3 545629
650 4 $a Wildlife conservation. $2 fast $3 542165
650 4 $a Field study. $3 3546147
650 4 $a Vegetation. $3 3560047
650 4 $a Principal components analysis. $3 565921
650 4 $a Remote sensing. $3 535394
650 4 $a Neural networks. $3 677449
650 4 $a Sensors. $3 3549539
650 4 $a Classification. $3 595585
650 4 $a Support vector machines. $3 2058743
650 4 $a Variables. $3 3548259
650 4 $a Methods. $3 3560391
650 4 $a Algorithms. $3 536374
650 4 $a Environmental protection. $3 527617
653 $a Mangroves
653 $a Deforestation
653 $a Biodiversity
655 7 $a Electronic books. $2 lcsh $3 542853
690 $a 0366
690 $a 0404
690 $a 0528
690 $a 0284
690 $a 0478
690 $a 0799
710 2 $a ProQuest Information and Learning Co. $3 783688
710 2 $a The Chinese University of Hong Kong (Hong Kong). $3 1017547
773 0 $t Dissertations Abstracts International $g 83-01B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=28663006 $z click for full text (PQDT)